Spooky season is here, and right on cue, MTA is clawing its way out of the grave. But no matter how many brands want to bring it back to life, or worse, prop it up with MMM and testing, MTA doesn't need a resurrection. It needs a eulogy. Let's give it a proper one and leave it back in 2018, where it belongs.
If that doesn’t work, I have it on good authority a wooden stake will do.
Heritage: The birth of MTA
Google adwords launched October of 2000. Google’s revenue went from $86M in 2001 to $10.6B in 2006 (12,225% growth). Last click attribution was standard in these simple times. Desktop was all that existed. The customer journey essentially consisted of arrival only.
- Google Click Auction rolled out 2002
- Shopify launched its ecommerce platform in 2006
- iPhones came out in 2007
- Facebook Self Serve ads launched November 2007
- Stripe launched its payment platform in 2011
- Facebook ads took off ~2014
In 2000, eCommerce was less than 1% of US retail. By 2020, it grew to 16%.
The earliest MTA movers started in 2008. They theorized that if they could observe their customer’s path to purchase, they would be able to deduce which ads were most influential in that purchase decision.
By 2012, the customer journey grew complex enough that crediting the last click for every conversion became fundamentally flawed. MTA stepped in promising an elegant solution, building click-paths and assigning credit to each touchpoint on the purchase journey.
Golden Years: MTA in its heyday
In the golden years of MTA, from 2014 to around 2017, customer paths were still short and desktop dominated online activity. Third-party cookies were unrestricted in every major browser.
Some will tell you privacy killed MTA. Really, changes across the user privacy landscape were just calling the time of death. MTA was biased from day one; it's just that, early on, cheaper media made being wrong more affordable.
Facebook ad revenue went from $4.28 billion in 2012 to $17.08 billion in 2015 to $69.66 billion in 2019, with mobile hitting 69% of ad revenue by Q4 2014.
US ecommerce grew from 4.5% of retail in 2010 to 10.6% in 2019. Budgets were growing along with every digital channel they were being spent on. Inventory was so cheap, and the reporting looked so good, accuracy just didn’t matter much.
Final Moments: The Death of MTA
In 2018, MTA quietly moved into the great beyond. In fact, Measured publicly declared it dead when we launched our first incrementality measurement solutions.
Atlas, the MTA solution Meta paid Microsoft between $50-100 million to acquire, was shut down. Google removed the DoubleClick ID in May 2018. Within just 8 weeks, both walled gardens closed the last channels through which a third party could reconstruct a user-level path across their inventory. That is the point in time that MTA stopped being technically possible, independent of whether it was ever methodologically sound.
Google had acquired Adometry in May of 2014 and spent nearly two years rebuilding it. They relaunched it as “Attribution 360” in March of 2016. It was sunset years later with no public notice. By 2023, fewer than 3% of conversions in Google Ads used a multi-touch attribution model.
General Data Protection Regulation, GDPR took effect May 25, 2018. The regulation applied to Europe, but platforms chose to apply the restrictions globally rather than run two architectures.
Reflections: MTA creates noise, not a unique signal
Most MTA post-mortems fixate on cause of death. Given recent resurrection attempts, the better question is whether it ever really had a pulse to begin with. Spoiler: MTA was always missing pieces, and no amount of necromancy will make it whole.
- It cannot see anything it cannot tag
- Marketing channels like linear TV, audio, OOH, and direct mail can't be measured with MTA.
- Sales channels like retail, wholesale, and marketplaces like Amazon can't be measured with MTA.
- Non-media drivers like price, promotions, seasonality, and macroeconomic conditions aren't accounted for at all.
- Tracking limitations
- MTA needs to know it's the same person at every touchpoint. The reliable ways to confirm that (cookies, pixels, tracking permissions) keep disappearing. Less certainty in means less certainty out.
- Ecosystem boundaries
- Third-party MTA can stitch together a user journey on the open web, but goes mostly blind once a touchpoint crosses into a platform that restricts what it shares with outside tools.
- Platform-native DDA sees everything inside its own walls and nothing outside them.
- Different tools, same limitation, just from opposite sides of the wall. Even machine learning-powered MTA can't account for the full picture.
- Platforms grade their own homework
- Platform-native attribution has an obvious incentive to credit itself generously. When every platform does this, attributed totals routinely add up to more conversions than the business actually had, and the numbers never tie back to what finance sees.
- It overvalues bottom-of-funnel touchpoints
- MTA never solved how heavily it over-credits click-based channels and under-credits view-based ones.
- Attempts to fix this with weighting only introduce bias, not science.
- This bias compounds over time. Because MTA over-credits bottom-funnel channels, brands shift budget there, which generates more attributed conversions, which "proves" the shift was right. It's a feedback loop that slowly starves the upper funnel.
- It's short-term focused
- MTA can't track conversions that happen after a session or cookie window lapses. That's fine if you're selling $9.99 drop-ship t-shirts, but it falls apart the moment you're trying to build a brand.
- It conflates an impression with an impact
- MTA was never able to answer the question, "How many of these conversions would have happened anyway?"
- It can't model diminishing returns
- MTA tells you what got credit, not what happens if you spend 20% more. With no response curves or saturation insight, it can't guide budget planning or forecasting.
- It can't be validated on its own
- MTA has no built-in ground truth. Without experiments like geo holdouts to reveal incrementality, there's no way of knowing whether the model is right, which is exactly why MTA results shouldn't be calibrating MMM or informing testing strategy.
Final Thoughts: Building an MMM on MTA signal isn't just spooky. It's downright terrifying.
Feeding MTA conversions into an MMM is the data scientist's version of Sinister: you found the cursed tapes in the attic, and you pressed play anyway.
- MTA isn't incremental, so it's the wrong calibration target. Calibration should anchor MMM to causal ground truth, like geo holdouts. Anchoring to MTA calibrates the model to credit, not lift, which quietly bakes the "would have happened anyway" conversions right into your MMM.
- MMM needs a consistent signal. Change anything about how the MTA works (its weights, its logic, even a genuine improvement) and you've changed the signal the model was built on. That leaves you with two bad choices; an MTA frozen in time forever or an MMM built on shifting ground.
- MTA bakes in bias. Weighting touchpoints by funnel position is a judgment call dressed up as math. Building your MMM on top of it doesn't remove the bias. It just removes objectivity.
- MMM gives that bias a disguise. An MTA report at least admits what it is: touch-based attribution. Run the same numbers through an MMM and they come back with a decomposition chart, confidence intervals, and a data science team attached. It looks like independent validation. It's the same biased number wearing a lab coat.
- It resurrects the very problem MMM was built to escape. MMM was developed to measure advertising performance before the internet or digital marketing even existed. Its recent resurgence happened because identity-based tracking failed and computing power took off. Anchoring your calibration to deterministic, identity-based MTA inputs invites all that signal loss right back in through the front door. You don't exorcise a ghost by moving it into a new house.
That's a lot of ghosts to invite into one model. At some point, you have to stop trying to raise the dead and focus on measuring the living.
MTA had its time. Let it rest in peace, and calibrate your MMM with something that actually measures causality: experiments.
Download the CMO's Guide to Causal Media Mix Modeling
A Farewell Musing
Here lies MTA, it’s last chapter is done,
They tweaked it for years, but the job stayed undone.
It traced what it could and lived up to its name,
But bias was baked in long before privacy came.
So stop haunting users wherever they scroll
Multi touch took the credit, without delivering the goal
For incrementality, here’s the best test:
Measure what caused it, let MTA rest.